On the Edge of AI for Wednesday, May 6: Anthropic grabs SpaceX GPUs, OpenAI rewrites data center networking, and a study says AI makes you dumber. Let's get into it.
Anthropic, the lab behind Claude, is taking over the full computing capacity of SpaceX's Colossus-1 data center [2]. More than 220,000 NVIDIA GPUs. Over 300 megawatts of power. The facility is expected to come online within a month. Alongside the hardware grab, the company is doubling rate limits for Claude Code and significantly raising API limits for Opus models. This isn't a small capacity add. It's a wholesale takeover of one of the largest single-site GPU clusters on the planet. The scale alone shifts the math for anyone planning a training run in the next twelve months. The signal: compute access is now a geopolitical asset, not just a procurement line item. My read: Anthropic isn't just buying chips. It's buying time. The next model cycle will demand more FLOPs than anyone has publicly committed to, and this deal locks down a massive slice before competitors can bid.
Different beat. OpenAI, the lab behind ChatGPT, has built a new open-source networking protocol called MRC with AMD, Broadcom, Intel, Microsoft, and NVIDIA [1]. The protocol sends data across hundreds of paths simultaneously between GPUs, bypassing the traditional three or four switch layers that have constrained cluster scaling for years. Instead of those layers, MRC needs only two to connect over 100,000 GPUs. It cuts both power and costs. The protocol is already running on OpenAI's Stargate supercomputer. What this changes: the bottleneck for scaling AI clusters isn't just chips anymore. It's how you wire them together. OpenAI is effectively standardizing the plumbing for the next generation of data centers.
Now. A new study suggests that using AI for just ten minutes can degrade your ability to think and problem-solve independently [4]. Cognitive offloading. People who lean on AI assistants show measurable drops in unassisted reasoning tasks shortly after, suggesting the brain treats the tool like a crutch rather than a collaborator. The angle: convenience has a tax, and it's measured in lost reasoning reps.
Last beat. Greg Brockman, OpenAI's co-founder, just shared his account of how Elon Musk left the company [3]. Cutthroat negotiations between startup founders are rarely shared so publicly, especially when a company becomes as world-changing as OpenAI and the stakes involve control of the most valuable technology platform of the decade. Which matters because: the origin story of the most valuable AI lab on earth is still being written in real time, and the fractures from those early days still shape the competitive map.
Hardware grabs, protocol wars, and the human cost of convenience. The infrastructure race is outpacing the model race.
That is the edge for today.